A self-adaptive modeling and control framework for HVAC systems considering user stochasticity

The modeling and control of HVAC systems are critical for energy savings and demand response programs. However, as part of the HVAC thermal system, user activities such as opening/closing windows and using appliances will impact the system thermal dynamics, which are stochastic in nature and cause modeling errors. Meanwhile, user's thermal preferences vary with time, weather, clothing, etc., which are also stochastic and make it difficult to design control strategies that can accurately optimize the energy costs and user comforts. In this paper, to better accommodate user stochasticity, a self-adaptive modeling and control framework is proposed. In the modeling part, a hybrid model integrating classical Equivalent Thermal Parameter (ETP) model with an advanced neural network TimesNet is proposed to capture both static building properties and dynamic user disturbances. In the control part, a two-stage control strategy combining Model Predictive Control (MPC) and Deep Q-Network (DQN) is designed to balance the control optimality and flexibility. Simulation results based on real-world building dataset demonstrate that the proposed hybrid model can achieve less than 1% of modeling error. The control strategy can achieve on average 13% energy savings while maintaining the temperature deviation within 0.5 °C from user preferences.